Insights · AI Consulting

What your board should ask before it approves an AI project.

More AI projects now reach the board before they reach production. Directors are asked to approve spend on work they cannot easily inspect, against risks that are hard to price. Eight questions that separate a project worth funding from one that is not, what a good board paper looks like, and the signs that a paper is not ready.

A year ago most AI spending sat inside an IT or innovation budget and never needed a board decision. That has changed. AI projects now touch customer data, staff roles and regulated processes, and the amounts involved are large enough to need approval. Directors in Australia are being asked to say yes or no to work that is new to most of them.

This piece is for the directors and executives on the other side of that paper. It sets out the questions worth asking before you approve an AI consulting or AI agent project, what a board paper should contain, and the warning signs that it needs more work. None of it requires the board to understand how a language model works.

Why this is a board question now.

Directors owe a duty of care and diligence under section 180 of the Corporations Act. Approving a material project without understanding what it does, what it risks and how it will be overseen sits badly with that duty, whatever the technology. Three things make AI projects harder than most:

  • The system keeps changing after approval. Models are updated, prompts are revised and the data shifts. What the board approved in March may behave differently in September.
  • Personal information is usually involved. From 10 December 2026, privacy policies must describe certain decisions made with the help of computer programs. Our piece on AI agents and the December privacy rule sets out what is covered.
  • Regulated entities carry extra obligations. APRA regulated entities must manage operational risk and material service providers under CPS 230, which has applied since 1 July 2025. Our CPS 230 guide covers how AI fits.

Eight questions to ask before you approve.

  • What work will this change, and how will we measure it? A named process, a baseline taken before the work starts, and a target. "Productivity" is not a measure. Hours returned to a team, cases handled without rework or days taken off a cycle are. Our guide to measuring an AI consulting engagement covers what good measures look like.
  • Is the data ready? Most AI projects that run late do so because the data was worse than anyone expected. Ask whether someone has looked at the actual data, not a description of it. An AI readiness assessment answers this before the money is committed.
  • Who is accountable when it is wrong? A named executive, not a vendor and not "the AI". The board should know who owns the system's outcomes and who can switch it off.
  • What can it do without a person? Reading, drafting and recommending carry different risks from sending, paying and deciding. Ask where the line sits and who approves anything past it. Our piece on what an AI agent should be allowed to do sets out the usual tiers.
  • Where does the data go? Which models and services are used, where they are hosted, and whether any personal information leaves Australia. Australian Privacy Principle 8 governs disclosure overseas. Our piece on where your AI data can live covers the options.
  • What will it cost to run, not only to build? Model usage, cloud services, monitoring and the people who keep it working. A paper that prices only the build has priced half the decision. Our piece on what AI consulting costs in Australia gives the usual ranges for the build.
  • How will we know it is still working in a year? An evaluation set of real cases, run on every change, and a report the board or a committee sees on a set rhythm. Our piece on what ongoing support should cover describes the minimum.
  • What do we own if we stop? The code, the prompts, the evaluation sets and the documentation, and whether the organisation can run the system without the firm that built it.

What a good board paper contains.

A good paper for an AI project is short. It answers the eight questions above on one or two pages, and puts the detail in an appendix. In practice that means:

  • The decision sought, the amount, and what is approved now versus later. Staged funding, with a go or no-go point after a working first release, is usually the safer shape.
  • The baseline and the target, with who measured the baseline and when.
  • The risks named, each with an owner: data quality, privacy, staff adoption, supplier dependence, and the cost of being wrong in the specific process.
  • The oversight plan: who reports to the board, how often, and what would trigger a pause.
  • The exit: what it costs to stop, and what the organisation keeps.

Signs the paper is not ready.

  • A demonstration in place of a plan. A good demonstration on sample data shows the vendor can build a demonstration. It says little about your data or your people.
  • No baseline. If nobody measured the process before, nobody will be able to say afterwards whether the project worked.
  • No named owner inside the organisation. If the accountable person is the vendor, the organisation has not yet decided it wants the system.
  • Every risk is rated low. AI work has known failure points. A paper that finds none has not looked.
  • One lump sum. A single approval for the whole programme, with no point at which the board sees evidence before committing the rest.

What the board should not do.

Boards do not need to choose the model, review prompts or approve each use case. That is management's job, and a board that tries to do it slows the work without reducing the risk. Nor is a blanket ban a safe answer: staff use AI tools whether or not they are approved, and an unmanaged tool is a larger risk than a governed one. The board's job is to set the appetite, ask the questions above and make sure someone is watching. Our AI governance consulting work starts from that division. Founders running smaller boards may find our notes on AI governance for founders useful too.

Why RUBIX.

RUBIX is an independent, Australian owned data and AI consultancy, founded in Melbourne in 2011. In 15 years we have delivered more than 450 projects for over 115 organisations, many of them in regulated industries where the board asks exactly these questions. We would rather a project be approved on evidence than on enthusiasm.

  • Papers that answer the questions. We write proposals that state the baseline, the named team, the assumptions behind the price, the running cost and what you will own.
  • Data first. We began as a data consultancy, so we look at the actual data before anyone asks the board for money.
  • Staged by default. We prefer a working first release and a go or no-go decision to one large approval, and our forward deployed engineers work inside your team so the knowledge stays with you.

If you are comparing firms for the work, our guide to comparing AI consulting proposals sets out what to check. If you would like help preparing a paper your board can approve with confidence, talk to us.

Frequently asked questions.

What should a board ask before approving an AI project?

What work it will change and how that will be measured, whether the data is ready, who is accountable when it is wrong, what it can do without a person, where the data goes, what it costs to run, how the board will know it is still working, and what the organisation owns if it stops.

Do directors need to understand how AI works to approve a project?

No. Directors need to understand what the system will do, what could go wrong, who is accountable and how it will be overseen. Choosing models and reviewing prompts is management's job.

Should a board approve an AI project in one decision or in stages?

Usually in stages. Approve a working first release with a go or no-go point, so the board sees evidence on real data before committing the rest of the budget.

Quick Summary: Before approving an AI project, a board should ask eight questions: what work it changes and how that is measured, whether the data is ready, who is accountable when it is wrong, what it can do without a person, where the data goes, what it costs to run as well as build, how the board will know it still works in a year, and what the organisation owns if it stops. A good paper answers them on two pages, names the risks with owners, and asks for staged funding. A demonstration with no baseline and no internal owner is not ready.

General information only, not legal, financial or regulatory advice. Current as at November 2026.